Python 将CNN从tf.layers重写为原始tf后性能不佳

在

tf.layers

模块的帮助下，我创建了一个简单的CNN，在MNIST数据库上对其进行训练

首先，我们加载数据：

import tensorflow as tf

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)

然后设置一些基本参数，建立并训练模型：

learning_rate = 0.01
training_epochs = 10
batch_size = 100

x = tf.placeholder(tf.float32, [None, 784], name='InputData')
y = tf.placeholder(tf.float32, [None, 10], name='LabelData')

with tf.name_scope('Model'):
    input_layer = tf.reshape(x, [-1, 28, 28, 1], name='InputReshaped')
    conv1 = tf.layers.conv2d(inputs=input_layer, filters=32, kernel_size=[4, 4], padding="same", activation=tf.nn.relu)
    pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2)
    dropout1 = tf.layers.dropout(inputs=pool1, rate=0.25)
    conv2 = tf.layers.conv2d(inputs=dropout1, filters=32, kernel_size=[4, 4], padding="same", activation=tf.nn.relu)
    pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2)
    dropout2 = tf.layers.dropout(inputs=pool2, rate=0.25)
    pool2_flat = tf.reshape(dropout2, [-1, 7 * 7 * 32])
    dense = tf.layers.dense(inputs=pool2_flat, units=256, activation=tf.nn.relu)
    dropout3 = tf.layers.dropout(inputs=dense, rate=0.5)
    pred = tf.layers.dense(inputs=dropout3, units=10)

with tf.name_scope('Loss'):
    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=pred, labels=y))

with tf.name_scope('SGD'):
    optimizer = tf.train.GradientDescentOptimizer(learning_rate)
    train_step = optimizer.minimize(loss)

with tf.name_scope('Accuracy'):
    acc = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
    acc = tf.reduce_mean(tf.cast(acc, tf.float32))

init = tf.global_variables_initializer()

with tf.Session() as sess:

    sess.run(init)

    for epoch in range(training_epochs):
        avg_cost = 0.
        avg_acc = 0.
        total_batch = int(mnist.train.num_examples/batch_size)
        for i in range(total_batch):
            batch_xs, batch_ys = mnist.train.next_batch(batch_size)
            _, c, ac = sess.run([train_step, loss, acc], feed_dict={x: batch_xs, y: batch_ys})
            avg_cost += c / total_batch
            avg_acc += ac / total_batch
        print("Epoch: {:04}, avg_cost = {:.9f}, avg_acc = {:.4f}".format(epoch + 1, avg_cost, avg_acc ))

    print("Optimization Finished!")

它工作正常，性能良好，并输出以下内容：

Epoch: 0001, avg_cost = 1.032827925, avg_acc = 0.7110
Epoch: 0002, avg_cost = 0.271804677, avg_acc = 0.9180
...
Epoch: 0010, avg_cost = 0.067859485, avg_acc = 0.9790
Optimization Finished!

但是，我想在不使用

tf.layers

的情况下重写模型。因此，我将

Model

块中的代码更改为以下内容-我认为其工作原理与前一个几乎相同：

def weight_variable(shape):
  initial = tf.truncated_normal(shape, stddev=0.1, mean = 0.1)
  return tf.Variable(initial)

def bias_variable(shape):
  initial = tf.constant(0.1, shape=shape)
  return tf.Variable(initial)

with tf.name_scope('Model'):
    with tf.name_scope('Input_L'):
        input_tsr = tf.reshape(x, [-1, 28, 28, 1], name='InputReshaped')
    with tf.name_scope('Conv1_L'):
        W_conv1 = weight_variable([4, 4, 1, 32])
        b_conv1 = bias_variable([32])
        conv1 = tf.add(tf.nn.conv2d(input_tsr, W_conv1, strides=[1, 1, 1, 1], padding='SAME'), b_conv1)
        h_conv1 = tf.nn.relu(conv1)
        h_pool1 = tf.nn.max_pool(h_conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
        dropout1 = tf.nn.dropout(h_pool1, 0.75)
    with tf.name_scope('Conv2_L'):
        W_conv2 = weight_variable([4, 4, 32, 32])
        b_conv2 = bias_variable([32])
        conv2 = tf.add(tf.nn.conv2d(dropout1, W_conv2, strides=[1, 1, 1, 1], padding='SAME'), b_conv2)
        h_conv2 = tf.nn.relu(conv2)
        h_pool2 = tf.nn.max_pool(h_conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
        dropout2 = tf.nn.dropout(h_pool2, 0.75)
    with tf.name_scope('Dense_L'):
        W_dense = weight_variable([7 * 7 * 32, 256])
        b_dense = bias_variable([256])
        flat_tsr = tf.reshape(dropout2, [-1, 7 * 7 * 32])
        dense = tf.add(tf.matmul(flat_tsr, W_dense), b_dense)
        h_dense =  tf.nn.relu(dense)
        dropout3 = tf.nn.dropout(h_dense, 0.5)
    with tf.name_scope('Output_L'):
        W_out = weight_variable([256, 10])
        b_out = bias_variable([10])
        pred = tf.add(tf.matmul(dropout3, W_out), b_out)

不幸的是，它的性能非常差，无法获得高于

0.12

的精度，我认为这意味着模型正在猜测正确的答案

Epoch: 0001, avg_cost = 22.226242821, avg_acc = 0.1106
Epoch: 0002, avg_cost = 2.301470806, avg_acc = 0.1123
...
Epoch: 0010, avg_cost = 2.301233784, avg_acc = 0.1123
Optimization Finished!

---

为什么第二种模式不能正确学习？你能指出第一个模型和第二个模型之间的区别在哪里（权重和偏差初始化除外）？

我认为文档中没有提到它，但是对于

tf.layers

子模块中的层，当提供

None

时，变量初始值设定项默认为

glorot\u uniform\u初始值设定项

---

如果你相应地替换了你的体重定义，你应该更接近你以前的结果。

事实上它解决了这个问题。我怀疑初始值是否会对性能产生如此大的影响。现在我认为初始值和低学习率的结合使得算法陷入了局部极小值。